AN AI-DRIVEN NETWORK SECURITY FRAMEWORK FOR INTELLIGENT THREAT DETECTION AND MITIGATION IN MODERN COMPUTER SYSTEMS

Abstract

Network intrusion detection research can overstate deployment readiness when preprocessing, repeated traffic patterns, or train-test overlap are not controlled. This research proposes a leakage-aware AI network security system on the basis of UNSW-NB15 benchmark and combines comparative machine learning, independent evaluation, Model-Agnostic Network security interpretation, and attack-informed mitigation suggestion. A total of 175,341 training records were used in the analysis and an 82,332-record prepared test partition (excluding 8,541 test records in which predictor vectors were duplicated in training) was used for primary evaluation. A stratified group-aware cross validation has been performed for comparing the logistic regression, decision tree, random forest and extremely randomized trees. Random forest had the highest mean cross-validated ROC-AUC (0.9908) and after tuning, it got the highest accuracy (0.9071), recall (0.9642), F1-score (0.9150), ROC-AUC (0.9820) and average precision (0.9848) on the clean test set. Permission analysis revealed contextual traffic characteristics and error analysis revealed that most attacks missed were Fuzzer traffic. A two stage response layer provides mitigative actions that are based on severity. The framework focuses on the transparent and reproducible evaluation, and the clear separation of detection evidence and unvalidated automated mitigation effectiveness.